Detection of Texture-less Occluded Objects Using Deep Convolutional Neural Networks

Nowadays, modern object classification algorithms can outsmart humans for classification of simple non-occluded objects. Machines, however, are still unable to accomplish complex tasks, such as classification and localization of occluded objects with precision, which human beings can easily perform. In order to achieve this task, convolutional neural networks based object detection algorithms have recently opened a new avenue towards the object detection with high precision. The convolutional neural networks require a large amount of training data to train system for classification and localization. However, the main objective of this thesis is to examine the convolutional neural network based object detection methods for the detection of texture-less occluded objects with limited amount of training data. The features extracted by convolutional layers contain more information as compared to hand engineered feature extraction methods; such as, SIFT and HOG. Convolutional layers extract features at different levels; initial layers extract information about edges while deeper layers extract more robust features that cover broader context of the image. The features extracted from deeper layers can identify texture-less occluded objects more precisely as compared to other traditional object detection methods, which includes edge based method and deform-able parts model using histogram of oriented gradient (HOG). Inspired from the success of convolutional neural network, this research has opted Single Shot Detector (SSD) and Faster Region based Convolution Network (Faster R-CNN) to accomplish the main objective. Mobilenet is the base i model in SSD; whereas, Inception is the base model in Faster R-CNN. SSD is superior than Faster R-CNN in terms of speed, but inferior in terms of average precision. A large number of training samples are required in SSD to develop fast run-time object detection. This research finds that SSD is not efficient for occluded object detection when input training data is limited. On the other hand, Faster R-CNN is comparatively slow in terms of speed, but average precision is significantly high as compared to other methods including SSD, edge based method, and deform-able parts model using HOG. This thesis concludes that Faster R-CNN has superior performance for occluded objects with limited training data. ii

Name thesis:degree_name

Master of Applied Science (MASc)

Level thesis:degree_level

Master's

Discipline thesis:degree_discipline

Engineering - Electronic Systems

Grantor dc:publisher

Faculty of Graduate Studies and Research, University of Regina

Year dc:date.issued

2018

Author dc:creator

• Irfan, Muhammad Faheem

Advisor dc:contributor.advisor

• Paranjape, Raman

Committee members dc:contributor.committeemember

• Al-Angabi, Irfan
• Wang, Zhanle

Language dc:language.iso

en

DOI

https://doi.org/10.82465/3865

OAI identifier oai:identifier

oai:uregina.scholaris.ca:10294/8533